Diagnosis selects the most probable explanation from a list of candidates the model already knows about. Abduction constructs the best complete assignment of all hidden variables — not selecting from a list, but finding the full account of the world that best fits everything observed. When the unexpected happens, diagnosis narrows down. Abduction opens up.
Three Modes of Inference
The three modes differ in which element of the inference triad is known and which is derived:
| Mode | Structure | What it produces | What it cannot produce |
|---|---|---|---|
| Deduction | Given rule + case → derive result All metals conduct. This is metal. Therefore it conducts. |
Necessary conclusions from given premises. Zero uncertainty if premises hold. | Anything not already contained in the premises. Deduction is truth-preserving, not truth-generating. |
| Induction | Given cases + results → derive rule These 1,000 metals all conducted. Therefore metals conduct. |
General rules from observed instances. Probabilistic — stronger with more cases. | Explanation of why the rule holds. Induction establishes that a pattern exists; it does not explain it. |
| Abduction | Given rule + result → hypothesize case Metals conduct. This conducted. Perhaps it is a metal. |
The best explanation for an observed result — the hypothesis that, if true, would make the observation least surprising. | Certainty. Abduction produces the most plausible hypothesis, not a guaranteed conclusion. It is explicitly tentative. |
The asymmetry between the three is precise: deduction and induction both work with observations that are already categorized. Abduction works when the categorization itself is uncertain — when you observe an effect and need to hypothesize what kind of thing could have produced it. It is the reasoning mode that precedes and enables the other two: before you can deduce consequences from a cause or induce a rule from cases, you need a hypothesis about what is going on.
Prediction (forward inference) gives you: given what I know about causes, what effects should I expect? Diagnosis (backward inference) gives you: given the effect I observed, which known cause most probably produced it? Both assume the cause is already in the model — already a node in the graph with its causal relationships encoded. Abduction operates when that assumption fails: when the observation is not well-explained by any existing cause in the model, and when the task is to construct the explanation that fits rather than select the best fitting existing one.
What Diagnosis Cannot Do
Diagnostic reasoning works backward from observed effect to most probable cause, running Bayes’ theorem over the encoded causal structure. The full treatment is on the Diagnostic Reasoning page. Three conditions it cannot handle:
- The observation doesn’t fit any known cause well — diagnosis ranks existing hypotheses; it cannot generate new ones.
- Multiple anomalies arrive simultaneously — diagnosis handles each observation; it cannot find a single explanation that accounts for all of them.
- The cause is partially outside the model — diagnosis is bounded by what was encoded; abduction can reason about what was omitted.
Abduction in a Causal Model
Standard diagnostic inference computes, for each possible cause X, the marginal posterior P(X | evidence). It answers: how probable is each candidate cause, considered individually? The results can be read independently — X is 78% probable, Y is 12% probable, Z is 6% probable — and the top-ranked candidate is the diagnosis.
MPE abduction computes something different: the single complete assignment of values to all hidden variables — every node in the graph that is not directly observed — that jointly maximises P(assignment | evidence). It answers: what is the most probable complete state of the world, given everything observed?
The distinction matters because marginal posteriors can be misleading when variables are dependent. A variable may have a high marginal posterior in isolation, but when combined with the jointly most probable value of an adjacent variable, it becomes inconsistent. MPE produces a coherent, mutually consistent account — not a collection of independently probable values that may contradict each other when assembled.
A claims model has two hidden variables: Fraud (true/false) and Claimant_distress (high/low). Both are causes of Claim_amount (elevated). Marginal diagnosis says: Fraud is 65% probable; Claimant_distress is 70% probable. These look like independent findings. But the MPE computation finds that the joint assignment (Fraud=false, Claimant_distress=high) has higher posterior probability than (Fraud=true, Claimant_distress=high) — because the two causes compete to explain the elevated claim amount, and once distress is assigned as the explanation, the evidence for fraud diminishes. The MPE answer: the best complete explanation is a distressed but legitimate claimant. The marginal answers suggested both fraud and distress were probable — but not jointly, and not as a coherent account.
The role of the prior in abductive inference
Abduction is explicitly prior-dependent. The “best” explanation is best relative to the model's prior beliefs and likelihood functions — which encode the domain experts' knowledge about how causes produce effects. Two models with different priors may produce different MPE answers from the same evidence. This is not a deficiency — it is the correct formalization of the fact that expert background knowledge shapes which explanations are considered plausible. Making the prior explicit and auditable is precisely the advantage of the causal model over informal expert judgment: the assumptions that drive the abductive conclusion are visible and challengeable.
Where It Applies
| Domain | The abductive question | Why diagnosis alone is insufficient |
|---|---|---|
| Fraud investigation | What is the complete account of this claim or transaction pattern — the full assignment of intent, circumstances, and intermediary behavior — that best explains the observed anomalies? | Fraud patterns are designed to defeat known diagnostic hypotheses. Novel schemes will not match any existing cause category well. The MPE computation finds the coherent account; marginal diagnosis finds the closest known pattern, which may be misleading. |
| Incident investigation | What is the full causal configuration — human factors, system states, environmental conditions, management decisions — that jointly produced the incident? | Major incidents are almost always multi-causal. Marginal diagnosis identifies the most probable single cause. MPE abduction finds the joint assignment that accounts for the full evidence — including why several controls failed simultaneously, which marginal diagnosis treats as independent questions. |
| Strategic anomaly explanation | Why did the market, customer base, or risk portfolio behave differently from the model's predictions — and what is the underlying state of the world that would make the observed divergence expected? | When reality diverges from forecast, the cause is often a structural shift — a change in a latent variable the model was not tracking. Abduction generates the hypothesis about what changed; the model can then be updated to incorporate it. |
| Expert knowledge elicitation | When a domain expert says “something is off but I can't say exactly what,” the abductive task is to formalize what their pattern recognition is pointing at. | Expert intuition is often abductive — a sense that the full picture doesn't fit a known explanation. The causal model provides the structured environment in which the expert's implicit hypothesis can be made explicit, tested, and encoded — converting tacit abduction into an auditable model component. |
Your experts reason abductively every time they encounter something that doesn't quite fit. The question is whether that reasoning is captured in a model the organization owns, or lost when they leave.
info@rung3.ai
Peirce, C.S., 1878, “Deduction, Induction, and Hypothesis,” Popular Science Monthly 13 · Pearl, J., 1988, Probabilistic Reasoning in Intelligent Systems, Morgan Kaufmann · Josephson, J.R. & Josephson, S.G. (eds.), 1994, Abductive Inference, Cambridge University Press · Peng, Y. & Reggia, J.A., 1990, Abductive Inference Models for Diagnostic Problem-Solving, Springer · Feldbacher-Escamilla, C.J. & Gebharter, A., 2018, “Modeling Creative Abduction Bayesian Style,” European Journal for Philosophy of Science — on positing new latent variables to explain correlated dispositions, the complementary move to MPE (selective) abduction discussed above.